Add files using upload-large-folder tool

2c07d60 verified 2 months ago

7.18 kB

	#!/usr/bin/env python3
	"""
	WebICoder v3 — Quick Start Example
	Generate HTML websites from natural language prompts using MLX on Apple Silicon.

	⚠️ MANDATORY: This script implements all 5 required rules for correct output.
	See README.md for full documentation.

	Usage:
	python example.py "Create a landing page for a coffee shop"
	python example.py --interactive
	"""

	import sys
	import re

	from mlx_lm import load, stream_generate
	from mlx_lm.sample_utils import make_sampler, make_logits_processors


	# ─── Configuration ──────────────────────────────────────────────────────────
	MODEL_PATH = "." # Current directory (the model repo)

	# RULE 1: System prompt + Alpaca format (### Instruction / ### Response)
	SYSTEM_PROMPT = (
	"You are WebICoder, an expert frontend web developer specializing in premium, "
	"Apple-inspired design. You create stunning websites using only HTML, CSS, and "
	"vanilla JavaScript. Your designs feature: minimalist layouts, elegant typography, "
	"smooth animations, glassmorphism effects, generous whitespace, and a refined "
	"color palette. You always produce complete, production-ready code."
	)

	# RULE 2: Stop sequences — MANDATORY to prevent infinite loops
	STOP_SEQUENCES = ["</html>", "### Instruction:", "You are Deepcoder", "You are WebICoder"]

	# RULE 4: Low temperature — MANDATORY for coherent HTML
	DEFAULT_TEMP = 0.4
	DEFAULT_MAX_TOKENS = 4096


	# ─── RULE 1: Prompt Formatting (MANDATORY) ──────────────────────────────────

	def format_prompt(user_input: str) -> str:
	"""
	MANDATORY: Format user input into the model's training prompt format.

	The model was trained with Alpaca-style prompts. Sending raw text
	without this formatting will produce garbage output.
	"""
	return f"{SYSTEM_PROMPT}\n\n### Instruction:\n{user_input}\n\n### Response:\n"


	# ─── RULE 5: Post-Processing (MANDATORY) ────────────────────────────────────

	def clean_html(text: str) -> str:
	"""
	MANDATORY: Extract clean HTML from model output.

	The model may leak training artifacts (system prompt, instruction markers).
	This function strips them and returns only valid HTML.
	"""
	# Remove system prompt leaks
	for pattern in [
	r"You are (?:Deep\|Web[iI])coder.?production-ready code\.\n",
	r"### Instruction:.*",
	r"### Response:\s*",
	]:
	text = re.sub(pattern, "", text, flags=re.DOTALL)

	# Extract complete HTML document
	html_match = re.search(r"(<(?:!DOCTYPE\s+html\|html)[\s\S]*?</html>)", text, re.IGNORECASE)
	if html_match:
	return html_match.group(1).strip()

	# Fallback: find any HTML content and wrap it
	html_start = re.search(r"<(?:!DOCTYPE\|html\|head\|body\|link)", text, re.IGNORECASE)
	if html_start:
	html = text[html_start.start():].strip()
	if not html.lower().startswith("<!doctype"):
	html = "<!DOCTYPE html>\n<html>\n" + html
	if "</html>" not in html.lower():
	html += "\n</html>"
	return html

	return text.strip()


	# ─── Generation ─────────────────────────────────────────────────────────────

	def generate_html(prompt: str, temperature: float = DEFAULT_TEMP, max_tokens: int = DEFAULT_MAX_TOKENS) -> str:
	"""
	Generate HTML from a natural language prompt.

	Implements all 5 mandatory rules:
	1. Prompt formatting (### Instruction / ### Response)
	2. Stop at </html>
	3. Repetition penalty (1.2, context=256)
	4. Low temperature (0.4)
	5. Post-processing (clean_html)
	"""
	print(f"[INFO] Loading model from: {MODEL_PATH}")
	model, tokenizer = load(MODEL_PATH)

	# RULE 1: Format the prompt
	formatted_prompt = format_prompt(prompt)

	# RULE 4: Low temperature sampler
	sampler = make_sampler(temp=temperature)

	# RULE 3: Repetition penalty — MANDATORY
	logits_processors = make_logits_processors(
	repetition_penalty=1.2,
	repetition_context_size=256,
	)

	print(f"[INFO] Generating (temp={temperature}, max_tokens={max_tokens}, rep_penalty=1.2)...")
	print("─" * 60)

	full_text = ""
	last_response = None

	for response in stream_generate(
	model, tokenizer,
	prompt=formatted_prompt,
	max_tokens=max_tokens,
	sampler=sampler,
	logits_processors=logits_processors, # RULE 3
	):
	last_response = response
	token_str = response.text
	full_text += token_str
	print(token_str, end="", flush=True)

	# RULE 2: Stop at </html> — MANDATORY
	should_stop = False
	for stop_seq in STOP_SEQUENCES:
	if stop_seq in full_text:
	idx = full_text.find(stop_seq)
	if stop_seq == "</html>":
	full_text = full_text[:idx + len(stop_seq)]
	else:
	full_text = full_text[:idx]
	should_stop = True
	break

	if should_stop or response.finish_reason is not None:
	break

	print("\n" + "─" * 60)
	if last_response:
	print(f"[INFO] Generated {last_response.generation_tokens} tokens at {last_response.generation_tps:.1f} tok/s")
	print(f"[INFO] Peak memory: {last_response.peak_memory:.2f} GB")

	# RULE 5: Clean the output — MANDATORY
	return clean_html(full_text)


	# ─── Main ────────────────────────────────────────────────────────────────────

	def main():
	if len(sys.argv) > 1 and sys.argv[1] != "--interactive":
	# Single prompt mode
	prompt = " ".join(sys.argv[1:])
	html = generate_html(prompt)

	output_file = "output.html"
	with open(output_file, "w") as f:
	f.write(html)
	print(f"\n[INFO] Saved to {output_file} ({len(html)} chars)")

	else:
	# Interactive mode
	print("=" * 60)
	print(" ⚡ WebICoder v3 — Interactive Mode")
	print(" Type a website description, press Enter to generate.")
	print(" Type 'quit' to exit.")
	print("=" * 60)

	while True:
	try:
	prompt = input("\n🌐 Describe your website: ").strip()
	if not prompt or prompt.lower() in ("quit", "exit", "q"):
	break

	html = generate_html(prompt)

	output_file = "output.html"
	with open(output_file, "w") as f:
	f.write(html)
	print(f"\n[INFO] Saved to {output_file} ({len(html)} chars)")

	except KeyboardInterrupt:
	print("\n[INFO] Bye!")
	break


	if __name__ == "__main__":
	main()